A DeadMod user liked almost all of a generated lyric video. Then three lyric lines disappeared.
That small failure ruined the result. A music video cannot be “mostly correct” when the audience knows every word.
Transcription has several truths
A transcription provider can return full text, phrase segments, and timestamps for each word. Those parts do not always agree.
Punctuation changes. Contractions split. A segment ends a fraction too early. A word lands between two phrases. If the player trusts segment text, words can vanish. If it groups every fixed number of words, the lyrics stop feeling like lyrics.
DeadMod now treats word timestamps as the timing source of truth. Provider segments are useful hints for line breaks, but they do not own the displayed text.
Each timed word goes into the segment that covers its start time. If no segment covers it, the word joins the nearest segment. The renderer rebuilds each lyric line from the timed words.
The user still needs control
Automatic timing will not be perfect for every vocal, accent, mix, or song structure.
I added manual correction for the bad section instead of asking the user to rebuild the song. The editor preserves known timings where possible and limits the amount of text changed at once.
Preview and export also share the same timing, fitting, and layout helpers. A correction that looks right in the editor must look the same in the finished video.
That was the real lesson. Creative software is judged by the part that breaks the final artifact. The first 95% makes a good demo. The last 5% makes a usable product.